Index: /trunk/modules/vci_mem_cache/caba/source/include/vci_mem_cache.h
===================================================================
--- /trunk/modules/vci_mem_cache/caba/source/include/vci_mem_cache.h (revision 662)
+++ /trunk/modules/vci_mem_cache/caba/source/include/vci_mem_cache.h (revision 663)
@@ -405,4 +405,5 @@
// instrumentation counters
uint32_t m_cpt_cycles; // Counter of cycles
+ uint32_t m_cpt_reset_count; // Cycle at which the counters were last reset
// Counters accessible in software (not yet but eventually)
@@ -449,5 +450,5 @@
uint32_t m_cpt_write_miss; // Number of MISS WRITE
uint32_t m_cpt_write_dirty; // Cumulated length for WRITE transactions
- uint32_t m_cpt_write_broadcast;// Number of BROADCAST INVAL because write
+ uint32_t m_cpt_write_broadcast;// Number of BROADCAST INVAL because of writes
uint32_t m_cpt_trt_rb; // Read blocked by a hit in trt
@@ -515,4 +516,5 @@
~VciMemCache();
+ void reset_counters();
void print_stats(bool activity_counters, bool stats);
void print_trace( size_t detailed = 0 );
Index: /trunk/modules/vci_mem_cache/caba/source/src/vci_mem_cache.cpp
===================================================================
--- /trunk/modules/vci_mem_cache/caba/source/src/vci_mem_cache.cpp (revision 662)
+++ /trunk/modules/vci_mem_cache/caba/source/src/vci_mem_cache.cpp (revision 663)
@@ -31,4 +31,5 @@
* alexandre.joannou@lip6.fr
*/
+
#include "../include/vci_mem_cache.h"
@@ -799,4 +800,47 @@
}
+
+ /////////////////////////////////////////
+ tmpl(void)::reset_counters()
+ /////////////////////////////////////////
+ {
+ m_cpt_reset_count = m_cpt_cycles;
+ m_cpt_read_local = 0;
+ m_cpt_read_remote = 0;
+ m_cpt_read_cost = 0;
+ m_cpt_write_local = 0;
+ m_cpt_write_remote = 0;
+ m_cpt_write_flits_local = 0;
+ m_cpt_write_flits_remote = 0;
+ m_cpt_write_cost = 0;
+ m_cpt_ll_local = 0;
+ m_cpt_ll_remote = 0;
+ m_cpt_ll_cost = 0;
+ m_cpt_sc_local = 0;
+ m_cpt_sc_remote = 0;
+ m_cpt_sc_cost = 0;
+ m_cpt_cas_local = 0;
+ m_cpt_cas_remote = 0;
+ m_cpt_cas_cost = 0;
+ m_cpt_update = 0;
+ m_cpt_update_local = 0;
+ m_cpt_update_remote = 0;
+ m_cpt_update_cost = 0;
+ m_cpt_minval = 0;
+ m_cpt_minval_local = 0;
+ m_cpt_minval_remote = 0;
+ m_cpt_minval_cost = 0;
+ m_cpt_binval = 0;
+ m_cpt_cleanup_local = 0;
+ m_cpt_cleanup_remote = 0;
+ m_cpt_cleanup_cost = 0;
+ m_cpt_read_miss = 0;
+ m_cpt_write_miss = 0;
+ m_cpt_write_dirty = 0;
+ m_cpt_trt_rb = 0;
+ m_cpt_trt_full = 0;
+ m_cpt_get = 0;
+ m_cpt_put = 0;
+ }
/////////////////////////////////////////
@@ -812,4 +856,5 @@
std::cout << "----------------------------------" << std::dec << std::endl;
std::cout
+ << "[000] COUNTERS RESET AT CYCLE = " << m_cpt_reset_count << std::endl
<< "[001] NUMBER OF CYCLES = " << m_cpt_cycles << std::endl
<< std::endl
@@ -941,5 +986,5 @@
delete [] m_debug_data;
- print_stats();
+ //print_stats();
}
@@ -1068,4 +1113,5 @@
// Activity counters
m_cpt_cycles = 0;
+ m_cpt_reset_count = 0;
m_cpt_read_local = 0;
m_cpt_read_remote = 0;
@@ -1206,5 +1252,4 @@
if (p_vci_tgt.cmdval)
{
-
#if DEBUG_MEMC_TGT_CMD
if (m_debug)
@@ -3705,6 +3750,4 @@
"MEMC ERROR in WRITE_BC_DIR_READ state: Bad DIR allocation");
- m_cpt_write_broadcast++;
-
// write enable signal for data buffer.
r_write_bc_data_we = true;
@@ -5509,5 +5552,5 @@
<< " address = " << std::hex << r_cleanup_nline.read()*m_words*4
<< " / way = " << std::dec << r_cleanup_way.read()
- << " / srcid = " << std::dec << r_cleanup_srcid.read()
+ << " / srcid = " << std::hex << r_cleanup_srcid.read()
<< std::endl;
#endif
@@ -6639,4 +6682,5 @@
//
m_cpt_binval++;
+ m_cpt_write_broadcast++;
//
Index: /trunk/platforms/almos-tsar-mipsel/top.desc
===================================================================
--- /trunk/platforms/almos-tsar-mipsel/top.desc (revision 662)
+++ /trunk/platforms/almos-tsar-mipsel/top.desc (revision 663)
@@ -23,5 +23,5 @@
uses = [
- Uses('caba:tsar_xbar_cluster',
+ Uses('caba:tsar_xbar_cluster_almos',
vci_data_width_int = vci_cell_size_int,
vci_data_width_ext = vci_cell_size_ext,
Index: /trunk/platforms/almos-tsar-mipsel/tsar_xbar_cluster/caba/metadata/tsar_xbar_cluster.sd
===================================================================
--- /trunk/platforms/almos-tsar-mipsel/tsar_xbar_cluster/caba/metadata/tsar_xbar_cluster.sd (revision 662)
+++ /trunk/platforms/almos-tsar-mipsel/tsar_xbar_cluster/caba/metadata/tsar_xbar_cluster.sd (revision 663)
@@ -2,5 +2,5 @@
# -*- python -*-
-Module('caba:tsar_xbar_cluster',
+Module('caba:tsar_xbar_cluster_almos',
classname = 'soclib::caba::TsarXbarCluster',
tmpl_parameters = [
Index: /trunk/platforms/tsar_generic_xbar/top.cpp
===================================================================
--- /trunk/platforms/tsar_generic_xbar/top.cpp (revision 662)
+++ /trunk/platforms/tsar_generic_xbar/top.cpp (revision 663)
@@ -91,13 +91,9 @@
#include "gdbserver.h"
#include "mapping_table.h"
+#include "alloc_elems.h"
#include "tsar_xbar_cluster.h"
-#include "alloc_elems.h"
-
-///////////////////////////////////////////////////
-// OS
-///////////////////////////////////////////////////
-
-//#define USE_ALMOS 1
-#define USE_GIET 1
+
+#define USE_ALMOS 1
+//#define USE_GIET
#ifdef USE_ALMOS
@@ -113,7 +109,16 @@
#endif
+#ifdef USE_ALMOS
+ #define PREFIX_OS "almos/"
+ #include "almos/hard_config.h"
+#endif
+#ifdef USE_GIET
+ #define PREFIX_OS "giet_vm/"
+#endif
+
///////////////////////////////////////////////////
// Parallelisation
///////////////////////////////////////////////////
+
#define USE_OPENMP 0
@@ -124,7 +129,7 @@
// cluster index (computed from x,y coordinates)
#ifdef USE_ALMOS
-# define cluster(x,y) (y + x * Y_MAX)
+ #define cluster(x,y) (y + x * Y_SIZE)
#else
-# define cluster(x,y) (y + (x << Y_WIDTH))
+ #define cluster(x,y) (y + (x << Y_WIDTH))
#endif
@@ -162,22 +167,7 @@
////////////////////////////////////////////////////////////
-// Main Hardware Parameters values
-//////////////////////i/////////////////////////////////////
-
-#ifdef USE_ALMOS
-#include "almos/hard_config.h"
-#define PREFIX_OS "almos/"
-#endif
-#ifdef USE_GIET
-#include "giet_vm/hard_config.h"
-#define PREFIX_OS "giet_vm/"
-#endif
-
-////////////////////////////////////////////////////////////
// Secondary Hardware Parameters
//////////////////////i/////////////////////////////////////
-#define XMAX X_SIZE
-#define YMAX Y_SIZE
#define XRAM_LATENCY 0
@@ -193,6 +183,6 @@
#ifdef USE_ALMOS
-#define FBUF_X_SIZE 512
-#define FBUF_Y_SIZE 512
+#define FBUF_X_SIZE 1024
+#define FBUF_Y_SIZE 1024
#endif
#ifdef USE_GIET
@@ -224,5 +214,5 @@
#ifdef USE_ALMOS
-#define soft_name PREFIX_OS"bootloader.bin",\
+#define soft_name PREFIX_OS"bootloader-tsar-mipsel.bin",\
PREFIX_OS"kernel-soclib.bin@0xbfc10000:D",\
PREFIX_OS"arch-info.bib@0xBFC08000:D"
@@ -236,6 +226,5 @@
//////////////////////i/////////////////////////////////////
-#define MAX_FROZEN_CYCLES 1000000
-
+#define MAX_FROZEN_CYCLES 100000000
@@ -249,11 +238,10 @@
#define MDMA_TGTID 2
#define MTTY_TGTID 3
-#define FBUF_TGTID 4
-#define BDEV_TGTID 5
-#define MNIC_TGTID 6
-#define BROM_TGTID 7
-#define CDMA_TGTID 8
-#define SIMH_TGTID 9
-
+#define BDEV_TGTID 4
+#define MNIC_TGTID 5
+#define BROM_TGTID 6
+#define CDMA_TGTID 7
+#define SIMH_TGTID 8
+#define FBUF_TGTID 9
@@ -307,7 +295,6 @@
// (Almos supports 32 bits physical addresses)
- #define CLUSTER_IO_INC (cluster_io_id * (0x80000000ULL / (XMAX * YMAX) * 2))
-
- #define MEMC_MAX_SIZE (0x80000000 / (XMAX * YMAX))
+ #define CLUSTER_IO_INC (cluster_io_id * (0x80000000ULL / (X_SIZE * Y_SIZE) * 2))
+ #define MEMC_MAX_SIZE (0x40000000 / (X_SIZE * Y_SIZE))
#define BROM_BASE 0x00BFC00000
@@ -315,5 +302,5 @@
#define MEMC_BASE 0x0000000000
- #define MEMC_SIZE min(0x02000000, MEMC_MAX_SIZE)
+ #define MEMC_SIZE min(0x04000000, MEMC_MAX_SIZE)
#define XICU_BASE MEMC_MAX_SIZE + (XICU_TGTID << 19)
@@ -325,22 +312,20 @@
#define BDEV_BASE MEMC_MAX_SIZE + (BDEV_TGTID << 19) + (CLUSTER_IO_INC)
#define BDEV_SIZE 0x0000001000 // 4 Kbytes
-
+
#define MTTY_BASE MEMC_MAX_SIZE + (MTTY_TGTID << 19) + (CLUSTER_IO_INC)
#define MTTY_SIZE 0x0000001000 // 4 Kbytes
-
+
#define FBUF_BASE MEMC_MAX_SIZE + (FBUF_TGTID << 19) + (CLUSTER_IO_INC)
#define FBUF_SIZE (FBUF_X_SIZE * FBUF_Y_SIZE * 2) // Should be 0x80000
-
+
#define MNIC_BASE MEMC_MAX_SIZE + (MNIC_TGTID << 19) + (CLUSTER_IO_INC)
#define MNIC_SIZE 0x0000080000
-
+
#define CDMA_BASE MEMC_MAX_SIZE + (CDMA_TGTID << 19) + (CLUSTER_IO_INC)
#define CDMA_SIZE (0x0000004000 * NB_CMA_CHANNELS)
-
+
#define SIMH_BASE MEMC_MAX_SIZE + (SIMH_TGTID << 19) + (CLUSTER_IO_INC)
#define SIMH_SIZE 0x0000001000
-
-#endif
-
+#endif
bool stop_called = false;
@@ -354,20 +339,25 @@
#ifdef USE_GIET
- char soft_name[256] = soft_pathname; // pathname to binary code
-#endif
- uint64_t ncycles = 0xFFFFFFFFFFFFFFFF; // simulated cycles
- char disk_name[256] = BDEV_IMAGE_NAME; // pathname to the disk image
- char nic_rx_name[256] = NIC_RX_NAME; // pathname to the rx packets file
- char nic_tx_name[256] = NIC_TX_NAME; // pathname to the tx packets file
- ssize_t threads_nr = 1; // simulator's threads number
- bool debug_ok = false; // trace activated
- size_t debug_period = 1; // trace period
- size_t debug_memc_id = 0; // index of memc to be traced
- size_t debug_proc_id = 0; // index of proc to be traced
- uint32_t debug_from = 0; // trace start cycle
- uint32_t frozen_cycles = MAX_FROZEN_CYCLES; // monitoring frozen processor
+ char soft_name[256] = soft_pathname; // pathname to binary code
+#endif
+ const int64_t max_cycles = 5000000; // Maximum number of cycles simulated in one sc_start call
+ int64_t ncycles = 0x7FFFFFFFFFFFFFFF; // simulated cycles
+ char disk_name[256] = BDEV_IMAGE_NAME; // pathname to the disk image
+ char nic_rx_name[256] = NIC_RX_NAME; // pathname to the rx packets file
+ char nic_tx_name[256] = NIC_TX_NAME; // pathname to the tx packets file
+ ssize_t threads_nr = 1; // simulator's threads number
+ bool debug_ok = false; // trace activated
+ size_t debug_period = 1; // trace period
+ size_t debug_memc_id = 0; // index of memc to be traced
+ size_t debug_proc_id = 0; // index of proc to be traced
+ int64_t debug_from = 0; // trace start cycle
+ int64_t frozen_cycles = MAX_FROZEN_CYCLES; // monitoring frozen processor
size_t cluster_io_id; // index of cluster containing IOs
- struct timeval t1,t2;
- uint64_t ms1,ms2;
+ int64_t reset_counters = -1;
+ int64_t dump_counters = -1;
+ bool do_reset_counters = false;
+ bool do_dump_counters = false;
+ struct timeval t1, t2;
+ uint64_t ms1, ms2;
////////////// command line arguments //////////////////////
@@ -378,5 +368,5 @@
if ((strcmp(argv[n], "-NCYCLES") == 0) && (n + 1 < argc))
{
- ncycles = (uint64_t) strtol(argv[n + 1], NULL, 0);
+ ncycles = (int64_t) strtol(argv[n + 1], NULL, 0);
}
else if ((strcmp(argv[n], "-SOFT") == 0) && (n + 1 < argc))
@@ -396,18 +386,17 @@
{
debug_ok = true;
- debug_from = (uint32_t) strtol(argv[n + 1], NULL, 0);
+ debug_from = (int64_t) strtol(argv[n + 1], NULL, 0);
}
else if ((strcmp(argv[n], "-MEMCID") == 0) && (n + 1 < argc))
{
debug_memc_id = (size_t) strtol(argv[n + 1], NULL, 0);
-
-#ifdef USE_ALMOS
- assert((debug_memc_id < (XMAX * YMAX)) &&
- "debug_memc_id larger than XMAX * YMAX" );
+#ifdef USE_ALMOS
+ assert((debug_memc_id < (X_SIZE * Y_SIZE)) &&
+ "debug_memc_id larger than X_SIZE * Y_SIZE" );
#else
size_t x = debug_memc_id >> Y_WIDTH;
size_t y = debug_memc_id & ((1<> (vci_address_width - x_width - y_width);
+
/////////////////////
@@ -578,7 +577,7 @@
0x00FF800000);
- for (size_t x = 0; x < XMAX; x++)
+ for (size_t x = 0; x < X_SIZE; x++)
{
- for (size_t y = 0; y < YMAX; y++)
+ for (size_t y = 0; y < Y_SIZE; y++)
{
sc_uint offset;
@@ -628,7 +627,7 @@
0xFFFF000000ULL);
- for (size_t x = 0; x < XMAX; x++)
+ for (size_t x = 0; x < X_SIZE; x++)
{
- for (size_t y = 0; y < YMAX ; y++)
+ for (size_t y = 0; y < Y_SIZE ; y++)
{
@@ -655,31 +654,31 @@
// Horizontal inter-clusters DSPIN signals
DspinSignals*** signal_dspin_h_cmd_inc =
- alloc_elems >("signal_dspin_h_cmd_inc", XMAX-1, YMAX, 3);
+ alloc_elems >("signal_dspin_h_cmd_inc", X_SIZE-1, Y_SIZE, 3);
DspinSignals*** signal_dspin_h_cmd_dec =
- alloc_elems >("signal_dspin_h_cmd_dec", XMAX-1, YMAX, 3);
+ alloc_elems >("signal_dspin_h_cmd_dec", X_SIZE-1, Y_SIZE, 3);
DspinSignals*** signal_dspin_h_rsp_inc =
- alloc_elems >("signal_dspin_h_rsp_inc", XMAX-1, YMAX, 2);
+ alloc_elems >("signal_dspin_h_rsp_inc", X_SIZE-1, Y_SIZE, 2);
DspinSignals*** signal_dspin_h_rsp_dec =
- alloc_elems >("signal_dspin_h_rsp_dec", XMAX-1, YMAX, 2);
+ alloc_elems >("signal_dspin_h_rsp_dec", X_SIZE-1, Y_SIZE, 2);
// Vertical inter-clusters DSPIN signals
DspinSignals*** signal_dspin_v_cmd_inc =
- alloc_elems >("signal_dspin_v_cmd_inc", XMAX, YMAX-1, 3);
+ alloc_elems >("signal_dspin_v_cmd_inc", X_SIZE, Y_SIZE-1, 3);
DspinSignals*** signal_dspin_v_cmd_dec =
- alloc_elems >("signal_dspin_v_cmd_dec", XMAX, YMAX-1, 3);
+ alloc_elems >("signal_dspin_v_cmd_dec", X_SIZE, Y_SIZE-1, 3);
DspinSignals*** signal_dspin_v_rsp_inc =
- alloc_elems >("signal_dspin_v_rsp_inc", XMAX, YMAX-1, 2);
+ alloc_elems >("signal_dspin_v_rsp_inc", X_SIZE, Y_SIZE-1, 2);
DspinSignals*** signal_dspin_v_rsp_dec =
- alloc_elems >("signal_dspin_v_rsp_dec", XMAX, YMAX-1, 2);
+ alloc_elems >("signal_dspin_v_rsp_dec", X_SIZE, Y_SIZE-1, 2);
// Mesh boundaries DSPIN signals
DspinSignals**** signal_dspin_false_cmd_in =
- alloc_elems >("signal_dspin_false_cmd_in" , XMAX, YMAX, 4, 3);
+ alloc_elems >("signal_dspin_false_cmd_in" , X_SIZE, Y_SIZE, 4, 3);
DspinSignals**** signal_dspin_false_cmd_out =
- alloc_elems >("signal_dspin_false_cmd_out", XMAX, YMAX, 4, 3);
+ alloc_elems >("signal_dspin_false_cmd_out", X_SIZE, Y_SIZE, 4, 3);
DspinSignals**** signal_dspin_false_rsp_in =
- alloc_elems >("signal_dspin_false_rsp_in" , XMAX, YMAX, 4, 2);
+ alloc_elems >("signal_dspin_false_rsp_in" , X_SIZE, Y_SIZE, 4, 2);
DspinSignals**** signal_dspin_false_rsp_out =
- alloc_elems >("signal_dspin_false_rsp_out", XMAX, YMAX, 4, 2);
+ alloc_elems >("signal_dspin_false_rsp_out", X_SIZE, Y_SIZE, 4, 2);
@@ -700,5 +699,5 @@
dspin_rsp_width,
vci_param_int,
- vci_param_ext>* clusters[XMAX][YMAX];
+ vci_param_ext>* clusters[X_SIZE][Y_SIZE];
#if USE_OPENMP
@@ -707,8 +706,8 @@
#pragma omp for
#endif
- for (size_t i = 0; i < (XMAX * YMAX); i++)
+ for (size_t i = 0; i < (X_SIZE * Y_SIZE); i++)
{
- size_t x = i / YMAX;
- size_t y = i % YMAX;
+ size_t x = i / Y_SIZE;
+ size_t y = i % Y_SIZE;
#if USE_OPENMP
@@ -768,5 +767,5 @@
loader,
frozen_cycles,
- debug_from ,
+ debug_from,
debug_ok and (cluster(x,y) == debug_memc_id),
debug_ok and (cluster(x,y) == debug_proc_id)
@@ -786,6 +785,6 @@
// Clock & RESET
- for (size_t x = 0; x < (XMAX); x++){
- for (size_t y = 0; y < YMAX; y++){
+ for (size_t x = 0; x < (X_SIZE); x++){
+ for (size_t y = 0; y < Y_SIZE; y++){
clusters[x][y]->p_clk (signal_clk);
clusters[x][y]->p_resetn (signal_resetn);
@@ -794,7 +793,7 @@
// Inter Clusters horizontal connections
- if (XMAX > 1){
- for (size_t x = 0; x < (XMAX-1); x++){
- for (size_t y = 0; y < YMAX; y++){
+ if (X_SIZE > 1){
+ for (size_t x = 0; x < (X_SIZE-1); x++){
+ for (size_t y = 0; y < Y_SIZE; y++){
for (size_t k = 0; k < 3; k++){
clusters[x][y]->p_cmd_out[EAST][k] (signal_dspin_h_cmd_inc[x][y][k]);
@@ -816,7 +815,7 @@
// Inter Clusters vertical connections
- if (YMAX > 1) {
- for (size_t y = 0; y < (YMAX-1); y++){
- for (size_t x = 0; x < XMAX; x++){
+ if (Y_SIZE > 1) {
+ for (size_t y = 0; y < (Y_SIZE-1); y++){
+ for (size_t x = 0; x < X_SIZE; x++){
for (size_t k = 0; k < 3; k++){
clusters[x][y]->p_cmd_out[NORTH][k] (signal_dspin_v_cmd_inc[x][y][k]);
@@ -838,5 +837,5 @@
// East & West boundary cluster connections
- for (size_t y = 0; y < YMAX; y++)
+ for (size_t y = 0; y < Y_SIZE; y++)
{
for (size_t k = 0; k < 3; k++)
@@ -844,6 +843,6 @@
clusters[0][y]->p_cmd_in[WEST][k] (signal_dspin_false_cmd_in[0][y][WEST][k]);
clusters[0][y]->p_cmd_out[WEST][k] (signal_dspin_false_cmd_out[0][y][WEST][k]);
- clusters[XMAX-1][y]->p_cmd_in[EAST][k] (signal_dspin_false_cmd_in[XMAX-1][y][EAST][k]);
- clusters[XMAX-1][y]->p_cmd_out[EAST][k] (signal_dspin_false_cmd_out[XMAX-1][y][EAST][k]);
+ clusters[X_SIZE-1][y]->p_cmd_in[EAST][k] (signal_dspin_false_cmd_in[X_SIZE-1][y][EAST][k]);
+ clusters[X_SIZE-1][y]->p_cmd_out[EAST][k] (signal_dspin_false_cmd_out[X_SIZE-1][y][EAST][k]);
}
@@ -852,11 +851,11 @@
clusters[0][y]->p_rsp_in[WEST][k] (signal_dspin_false_rsp_in[0][y][WEST][k]);
clusters[0][y]->p_rsp_out[WEST][k] (signal_dspin_false_rsp_out[0][y][WEST][k]);
- clusters[XMAX-1][y]->p_rsp_in[EAST][k] (signal_dspin_false_rsp_in[XMAX-1][y][EAST][k]);
- clusters[XMAX-1][y]->p_rsp_out[EAST][k] (signal_dspin_false_rsp_out[XMAX-1][y][EAST][k]);
+ clusters[X_SIZE-1][y]->p_rsp_in[EAST][k] (signal_dspin_false_rsp_in[X_SIZE-1][y][EAST][k]);
+ clusters[X_SIZE-1][y]->p_rsp_out[EAST][k] (signal_dspin_false_rsp_out[X_SIZE-1][y][EAST][k]);
}
}
// North & South boundary clusters connections
- for (size_t x = 0; x < XMAX; x++)
+ for (size_t x = 0; x < X_SIZE; x++)
{
for (size_t k = 0; k < 3; k++)
@@ -864,6 +863,6 @@
clusters[x][0]->p_cmd_in[SOUTH][k] (signal_dspin_false_cmd_in[x][0][SOUTH][k]);
clusters[x][0]->p_cmd_out[SOUTH][k] (signal_dspin_false_cmd_out[x][0][SOUTH][k]);
- clusters[x][YMAX-1]->p_cmd_in[NORTH][k] (signal_dspin_false_cmd_in[x][YMAX-1][NORTH][k]);
- clusters[x][YMAX-1]->p_cmd_out[NORTH][k] (signal_dspin_false_cmd_out[x][YMAX-1][NORTH][k]);
+ clusters[x][Y_SIZE-1]->p_cmd_in[NORTH][k] (signal_dspin_false_cmd_in[x][Y_SIZE-1][NORTH][k]);
+ clusters[x][Y_SIZE-1]->p_cmd_out[NORTH][k] (signal_dspin_false_cmd_out[x][Y_SIZE-1][NORTH][k]);
}
@@ -872,6 +871,6 @@
clusters[x][0]->p_rsp_in[SOUTH][k] (signal_dspin_false_rsp_in[x][0][SOUTH][k]);
clusters[x][0]->p_rsp_out[SOUTH][k] (signal_dspin_false_rsp_out[x][0][SOUTH][k]);
- clusters[x][YMAX-1]->p_rsp_in[NORTH][k] (signal_dspin_false_rsp_in[x][YMAX-1][NORTH][k]);
- clusters[x][YMAX-1]->p_rsp_out[NORTH][k] (signal_dspin_false_rsp_out[x][YMAX-1][NORTH][k]);
+ clusters[x][Y_SIZE-1]->p_rsp_in[NORTH][k] (signal_dspin_false_rsp_in[x][Y_SIZE-1][NORTH][k]);
+ clusters[x][Y_SIZE-1]->p_rsp_out[NORTH][k] (signal_dspin_false_rsp_out[x][Y_SIZE-1][NORTH][k]);
}
}
@@ -888,6 +887,6 @@
// network boundaries signals
- for (size_t x = 0; x < XMAX ; x++){
- for (size_t y = 0; y < YMAX ; y++){
+ for (size_t x = 0; x < X_SIZE ; x++){
+ for (size_t y = 0; y < Y_SIZE ; y++){
for (size_t a = 0; a < 4; a++){
for (size_t k = 0; k < 3; k++){
@@ -911,143 +910,212 @@
signal_resetn = true;
- if (gettimeofday(&t1, NULL) != 0)
- {
- perror("gettimeofday");
- return EXIT_FAILURE;
- }
-
- for (uint64_t n = 1; n < ncycles && !stop_called; n++)
- {
- // Monitor a specific address for L1 & L2 caches
- //clusters[0][0]->proc[0]->cache_monitor(0x800002c000ULL);
- //clusters[1][0]->memc->copies_monitor(0x800002C000ULL);
-
- if( (n % 5000000) == 0)
+ if (debug_ok) {
+ #if USE_OPENMP
+ assert(false && "OPEN MP should not be used with debug because of its traces");
+ #endif
+
+ if (gettimeofday(&t1, NULL) != 0) {
+ perror("gettimeofday");
+ return EXIT_FAILURE;
+ }
+
+ for (int64_t n = 1; n < ncycles && !stop_called; n++)
{
-
- if (gettimeofday(&t2, NULL) != 0)
- {
+ // Monitor a specific address for L1 & L2 caches
+ //clusters[0][0]->proc[0]->cache_monitor(0x800002c000ULL);
+ //clusters[1][0]->memc->copies_monitor(0x800002C000ULL);
+
+ if ((n % max_cycles) == 0)
+ {
+
+ if (gettimeofday(&t2, NULL) != 0)
+ {
+ perror("gettimeofday");
+ return EXIT_FAILURE;
+ }
+
+ ms1 = (uint64_t) t1.tv_sec * 1000ULL + (uint64_t) t1.tv_usec / 1000;
+ ms2 = (uint64_t) t2.tv_sec * 1000ULL + (uint64_t) t2.tv_usec / 1000;
+ std::cerr << "platform clock frequency " << (double) 5000000 / (double) (ms2 - ms1) << "Khz" << std::endl;
+
+ if (gettimeofday(&t1, NULL) != 0)
+ {
+ perror("gettimeofday");
+ return EXIT_FAILURE;
+ }
+ }
+
+
+ if (n == reset_counters) {
+ for (size_t x = 0; x < (X_SIZE); x++) {
+ for (size_t y = 0; y < Y_SIZE; y++) {
+ clusters[x][y]->memc->reset_counters();
+ }
+ }
+ }
+
+ if (n == dump_counters) {
+ for (size_t x = 0; x < (X_SIZE); x++) {
+ for (size_t y = 0; y < Y_SIZE; y++) {
+ clusters[x][y]->memc->print_stats(true, false);
+ }
+ }
+ }
+
+ if (debug_ok and (n > debug_from) and (n % debug_period == 0))
+ {
+ std::cout << "****************** cycle " << std::dec << n ;
+ std::cout << " ************************************************" << std::endl;
+
+ // trace proc[debug_proc_id]
+ size_t l = debug_proc_id % NB_PROCS_MAX ;
+ size_t y = (debug_proc_id / NB_PROCS_MAX) % Y_SIZE ;
+ size_t x = debug_proc_id / (Y_SIZE * NB_PROCS_MAX) ;
+
+ std::ostringstream proc_signame;
+ proc_signame << "[SIG]PROC_" << x << "_" << y << "_" << l ;
+ std::ostringstream p2m_signame;
+ p2m_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " P2M" ;
+ std::ostringstream m2p_signame;
+ m2p_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " M2P" ;
+ std::ostringstream p_cmd_signame;
+ p_cmd_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " CMD" ;
+ std::ostringstream p_rsp_signame;
+ p_rsp_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " RSP" ;
+
+ //clusters[x][y]->wi_proc[l]->print_trace();
+ //clusters[x][y]->signal_vci_ini_proc[l].print_trace(proc_signame.str());
+ //clusters[x][y]->signal_dspin_p2m_proc[l].print_trace(p2m_signame.str());
+ //clusters[x][y]->signal_dspin_m2p_proc[l].print_trace(m2p_signame.str());
+ //clusters[x][y]->signal_dspin_cmd_proc_i[l].print_trace(p_cmd_signame.str());
+ //clusters[x][y]->signal_dspin_rsp_proc_i[l].print_trace(p_rsp_signame.str());
+
+ //clusters[x][y]->xbar_rsp_d->print_trace();
+ //clusters[x][y]->xbar_cmd_d->print_trace();
+ //clusters[x][y]->signal_dspin_cmd_l2g_d.print_trace("[SIG]L2G CMD");
+ //clusters[x][y]->signal_dspin_cmd_g2l_d.print_trace("[SIG]G2L CMD");
+ //clusters[x][y]->signal_dspin_rsp_l2g_d.print_trace("[SIG]L2G RSP");
+ //clusters[x][y]->signal_dspin_rsp_g2l_d.print_trace("[SIG]G2L RSP");
+
+ // trace memc[debug_memc_id]
+ x = debug_memc_id / Y_SIZE;
+ y = debug_memc_id % Y_SIZE;
+
+ std::ostringstream smemc;
+ smemc << "[SIG]MEMC_" << x << "_" << y;
+ std::ostringstream sxram;
+ sxram << "[SIG]XRAM_" << x << "_" << y;
+ std::ostringstream sm2p;
+ sm2p << "[SIG]MEMC_" << x << "_" << y << " M2P" ;
+ std::ostringstream sp2m;
+ sp2m << "[SIG]MEMC_" << x << "_" << y << " P2M" ;
+ std::ostringstream m_cmd_signame;
+ m_cmd_signame << "[SIG]MEMC_" << x << "_" << y << " CMD" ;
+ std::ostringstream m_rsp_signame;
+ m_rsp_signame << "[SIG]MEMC_" << x << "_" << y << " RSP" ;
+
+ //clusters[x][y]->memc->print_trace();
+ //clusters[x][y]->wt_memc->print_trace();
+ //clusters[x][y]->signal_vci_tgt_memc.print_trace(smemc.str());
+ //clusters[x][y]->signal_vci_xram.print_trace(sxram.str());
+ //clusters[x][y]->signal_dspin_p2m_memc.print_trace(sp2m.str());
+ //clusters[x][y]->signal_dspin_m2p_memc.print_trace(sm2p.str());
+ //clusters[x][y]->signal_dspin_cmd_memc_t.print_trace(m_cmd_signame.str());
+ //clusters[x][y]->signal_dspin_rsp_memc_t.print_trace(m_rsp_signame.str());
+
+ // trace replicated peripherals
+ //clusters[1][1]->mdma->print_trace();
+ //clusters[1][1]->signal_vci_tgt_mdma.print_trace("[SIG]MDMA_TGT_1_1");
+ //clusters[1][1]->signal_vci_ini_mdma.print_trace("[SIG]MDMA_INI_1_1");
+
+
+ // trace external peripherals
+ //size_t io_x = cluster_io_id / Y_SIZE;
+ //size_t io_y = cluster_io_id % Y_SIZE;
+
+ //clusters[io_x][io_y]->brom->print_trace();
+ //clusters[io_x][io_y]->wt_brom->print_trace();
+ //clusters[io_x][io_y]->signal_vci_tgt_brom.print_trace("[SIG]BROM");
+ //clusters[io_x][io_y]->signal_dspin_cmd_brom_t.print_trace("[SIG]BROM CMD");
+ //clusters[io_x][io_y]->signal_dspin_rsp_brom_t.print_trace("[SIG]BROM RSP");
+
+ //clusters[io_x][io_y]->bdev->print_trace();
+ //clusters[io_x][io_y]->signal_vci_tgt_bdev.print_trace("[SIG]BDEV_TGT");
+ //clusters[io_x][io_y]->signal_vci_ini_bdev.print_trace("[SIG]BDEV_INI");
+ }
+
+ sc_start(sc_core::sc_time(1, SC_NS));
+ }
+ }
+ else {
+ int64_t n = 0;
+ while (!stop_called) {
+ if (gettimeofday(&t1, NULL) != 0) {
perror("gettimeofday");
return EXIT_FAILURE;
}
-
+ int64_t nb_cycles = max_cycles;
+ if (do_reset_counters) {
+ nb_cycles = min(nb_cycles, reset_counters - n);
+ }
+ if (do_dump_counters) {
+ nb_cycles = min(nb_cycles, dump_counters - n);
+ }
+
+ sc_start(sc_core::sc_time(nb_cycles, SC_NS));
+ n += nb_cycles;
+
+ if (do_reset_counters && n == reset_counters) {
+ // Reseting counters
+ for (size_t x = 0; x < (X_SIZE); x++) {
+ for (size_t y = 0; y < Y_SIZE; y++) {
+ clusters[x][y]->memc->reset_counters();
+ }
+ }
+ do_reset_counters = false;
+ }
+
+ if (do_dump_counters && n == dump_counters) {
+ // Dumping counters
+ for (size_t x = 0; x < (X_SIZE); x++) {
+ for (size_t y = 0; y < Y_SIZE; y++) {
+ clusters[x][y]->memc->print_stats(true, false);
+ }
+ }
+ do_dump_counters = false;
+ }
+
+
+ if (gettimeofday(&t2, NULL) != 0) {
+ perror("gettimeofday");
+ return EXIT_FAILURE;
+ }
ms1 = (uint64_t) t1.tv_sec * 1000ULL + (uint64_t) t1.tv_usec / 1000;
ms2 = (uint64_t) t2.tv_sec * 1000ULL + (uint64_t) t2.tv_usec / 1000;
- std::cerr << "platform clock frequency " << (double) 5000000 / (double) (ms2 - ms1) << "Khz" << std::endl;
-
- if (gettimeofday(&t1, NULL) != 0)
- {
- perror("gettimeofday");
- return EXIT_FAILURE;
- }
- }
-
- if (debug_ok and (n > debug_from) and (n % debug_period == 0))
- {
- std::cout << "****************** cycle " << std::dec << n ;
- std::cout << " ************************************************" << std::endl;
-
- // trace proc[debug_proc_id]
- size_t l = debug_proc_id % NB_PROCS_MAX ;
- size_t y = (debug_proc_id / NB_PROCS_MAX) % YMAX ;
- size_t x = (debug_proc_id / ((1 << y_width) * NB_PROCS_MAX)) % XMAX ;
-
- std::ostringstream proc_signame;
- proc_signame << "[SIG]PROC_" << x << "_" << y << "_" << l ;
- std::ostringstream p2m_signame;
- p2m_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " P2M" ;
- std::ostringstream m2p_signame;
- m2p_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " M2P" ;
- std::ostringstream p_cmd_signame;
- p_cmd_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " CMD" ;
- std::ostringstream p_rsp_signame;
- p_rsp_signame << "[SIG]PROC_" << x << "_" << y << "_" << l << " RSP" ;
-
- clusters[x][y]->proc[l]->print_trace();
- clusters[x][y]->wi_proc[l]->print_trace();
- clusters[x][y]->signal_vci_ini_proc[l].print_trace(proc_signame.str());
- clusters[x][y]->signal_dspin_p2m_proc[l].print_trace(p2m_signame.str());
- clusters[x][y]->signal_dspin_m2p_proc[l].print_trace(m2p_signame.str());
- clusters[x][y]->signal_dspin_cmd_proc_i[l].print_trace(p_cmd_signame.str());
- clusters[x][y]->signal_dspin_rsp_proc_i[l].print_trace(p_rsp_signame.str());
-
- clusters[x][y]->xbar_rsp_d->print_trace();
- clusters[x][y]->xbar_cmd_d->print_trace();
- clusters[x][y]->signal_dspin_cmd_l2g_d.print_trace("[SIG]L2G CMD");
- clusters[x][y]->signal_dspin_cmd_g2l_d.print_trace("[SIG]G2L CMD");
- clusters[x][y]->signal_dspin_rsp_l2g_d.print_trace("[SIG]L2G RSP");
- clusters[x][y]->signal_dspin_rsp_g2l_d.print_trace("[SIG]G2L RSP");
-
- // trace memc[debug_memc_id]
- x = debug_memc_id / YMAX;
- y = debug_memc_id % YMAX;
-
- std::ostringstream smemc;
- smemc << "[SIG]MEMC_" << x << "_" << y;
- std::ostringstream sxram;
- sxram << "[SIG]XRAM_" << x << "_" << y;
- std::ostringstream sm2p;
- sm2p << "[SIG]MEMC_" << x << "_" << y << " M2P" ;
- std::ostringstream sp2m;
- sp2m << "[SIG]MEMC_" << x << "_" << y << " P2M" ;
- std::ostringstream m_cmd_signame;
- m_cmd_signame << "[SIG]MEMC_" << x << "_" << y << " CMD" ;
- std::ostringstream m_rsp_signame;
- m_rsp_signame << "[SIG]MEMC_" << x << "_" << y << " RSP" ;
-
- clusters[x][y]->memc->print_trace();
- clusters[x][y]->wt_memc->print_trace();
- clusters[x][y]->signal_vci_tgt_memc.print_trace(smemc.str());
- clusters[x][y]->signal_vci_xram.print_trace(sxram.str());
- clusters[x][y]->signal_dspin_p2m_memc.print_trace(sp2m.str());
- clusters[x][y]->signal_dspin_m2p_memc.print_trace(sm2p.str());
- clusters[x][y]->signal_dspin_cmd_memc_t.print_trace(m_cmd_signame.str());
- clusters[x][y]->signal_dspin_rsp_memc_t.print_trace(m_rsp_signame.str());
-
- // trace replicated peripherals
-// clusters[1][1]->mdma->print_trace();
-// clusters[1][1]->signal_vci_tgt_mdma.print_trace("[SIG]MDMA_TGT_1_1");
-// clusters[1][1]->signal_vci_ini_mdma.print_trace("[SIG]MDMA_INI_1_1");
-
-
- // trace external peripherals
- size_t io_x = cluster_io_id >> y_width;
- size_t io_y = cluster_io_id & ((1 << x_width) - 1);
-
- clusters[io_x][io_y]->brom->print_trace();
- clusters[io_x][io_y]->wt_brom->print_trace();
- clusters[io_x][io_y]->signal_vci_tgt_brom.print_trace("[SIG]BROM");
- clusters[io_x][io_y]->signal_dspin_cmd_brom_t.print_trace("[SIG]BROM CMD");
- clusters[io_x][io_y]->signal_dspin_rsp_brom_t.print_trace("[SIG]BROM RSP");
-
-// clusters[io_x][io_y]->bdev->print_trace();
-// clusters[io_x][io_y]->signal_vci_tgt_bdev.print_trace("[SIG]BDEV_TGT");
-// clusters[io_x][io_y]->signal_vci_ini_bdev.print_trace("[SIG]BDEV_INI");
- }
-
- sc_start(sc_core::sc_time(1, SC_NS));
+ std::cerr << "cycle " << n << " platform clock frequency " << (double) nb_cycles / (double) (ms2 - ms1) << "Khz" << std::endl;
+ }
}
// Free memory
- for (size_t i = 0; i < (XMAX * YMAX); i++)
+ for (size_t i = 0; i < (X_SIZE * Y_SIZE); i++)
{
- size_t x = i / YMAX;
- size_t y = i % YMAX;
+ size_t x = i / Y_SIZE;
+ size_t y = i % Y_SIZE;
delete clusters[x][y];
}
- dealloc_elems >(signal_dspin_h_cmd_inc, XMAX - 1, YMAX, 3);
- dealloc_elems >(signal_dspin_h_cmd_dec, XMAX - 1, YMAX, 3);
- dealloc_elems >(signal_dspin_h_rsp_inc, XMAX - 1, YMAX, 2);
- dealloc_elems >(signal_dspin_h_rsp_dec, XMAX - 1, YMAX, 2);
- dealloc_elems >(signal_dspin_v_cmd_inc, XMAX, YMAX - 1, 3);
- dealloc_elems >(signal_dspin_v_cmd_dec, XMAX, YMAX - 1, 3);
- dealloc_elems >(signal_dspin_v_rsp_inc, XMAX, YMAX - 1, 2);
- dealloc_elems >(signal_dspin_v_rsp_dec, XMAX, YMAX - 1, 2);
- dealloc_elems >(signal_dspin_false_cmd_in, XMAX, YMAX, 4, 3);
- dealloc_elems >(signal_dspin_false_cmd_out, XMAX, YMAX, 4, 3);
- dealloc_elems >(signal_dspin_false_rsp_in, XMAX, YMAX, 4, 2);
- dealloc_elems >(signal_dspin_false_rsp_out, XMAX, YMAX, 4, 2);
+ dealloc_elems >(signal_dspin_h_cmd_inc, X_SIZE - 1, Y_SIZE, 3);
+ dealloc_elems >(signal_dspin_h_cmd_dec, X_SIZE - 1, Y_SIZE, 3);
+ dealloc_elems >(signal_dspin_h_rsp_inc, X_SIZE - 1, Y_SIZE, 2);
+ dealloc_elems >(signal_dspin_h_rsp_dec, X_SIZE - 1, Y_SIZE, 2);
+ dealloc_elems >(signal_dspin_v_cmd_inc, X_SIZE, Y_SIZE - 1, 3);
+ dealloc_elems >(signal_dspin_v_cmd_dec, X_SIZE, Y_SIZE - 1, 3);
+ dealloc_elems >(signal_dspin_v_rsp_inc, X_SIZE, Y_SIZE - 1, 2);
+ dealloc_elems >(signal_dspin_v_rsp_dec, X_SIZE, Y_SIZE - 1, 2);
+ dealloc_elems >(signal_dspin_false_cmd_in, X_SIZE, Y_SIZE, 4, 3);
+ dealloc_elems >(signal_dspin_false_cmd_out, X_SIZE, Y_SIZE, 4, 3);
+ dealloc_elems >(signal_dspin_false_rsp_in, X_SIZE, Y_SIZE, 4, 2);
+ dealloc_elems >(signal_dspin_false_rsp_out, X_SIZE, Y_SIZE, 4, 2);
return EXIT_SUCCESS;